Skip to content

Invoke v4.0.0rc2

Pre-release
Pre-release
Compare
Choose a tag to compare
@brandonrising brandonrising released this 14 Mar 22:12
· 4558 commits to main since this release
v4.0.0rc2
330e135

This is a Release Candidate. We strongly suggest backing up your database before testing to prevent data loss in case of any issues.

Please let us know if you run into anything unexpected.

What's New

RC1 supported the setting skip_model_hash in invokeai.yaml. In RC2, this is replaced by a more flexible setting hashing_algorithm.

New Model Manager

The model manager is rewritten in v4.0.0, both frontend and backend. This builds a foundation for future model architectures and brings some exciting new user-facing features:

  • Queued model downloads
  • Per-model preview images
  • Per-model default settings - choose a model’s default VAE, Scheduler, CFG Scale, etc.
  • User-defined trigger phrases for concepts/LoRAs and models - access by typing the < key in any prompt box
  • API key support for model marketplaces

Model Hashing

When you first run v4.0.0, it will take a while to start up as it does a one-time hash of all of your model files.

Do not panic.

Hashes provide a stable identifier for a model that is the same across every platform.

If you don’t care about this, you can disable the hashing using the hashing_algorithm setting in invokeai.yaml.

Canvas Improvements

The canvas uses a new method for compositing called gradient denoising. This eliminates the need for multiple “passes”, greatly reducing generation time on the canvas. This method also provides substantially improved visual coherence between the masked regions and the rest of the image.

The compositing settings on canvas allow for control over the gradient denoising process.

Major research & experimentation for this novel denoising implementation was led by @dunkeroni, and @blessedcoolant was responsible for managing integration into the canvas UI.

Bonus: Invoke Training (Beta)

As of v4.0.0, all references to training in the core invoke script now point to the Invoke Training Repo. Invoke Training offers a simple user interface for:

  • Textual Inversion Training
  • LoRA Training
  • Dreambooth Training
  • Pivotal Tuning Training

You can learn more about Invoke Training at https://github.com/invoke-ai/invoke-training

Minor UI/UX Enhancements:

  • Canvas Brush Size Scroll can now be inverted (Thanks @joshistoast!)
  • Images in the Canvas Staging Area can now be discarded individually (Thanks @joshistoast!)
  • Many small bug fixes and resolved papercuts

Installation and Upgrading

To install or upgrade to version 4.0, download the zip file from the release notes ("Assets" section), unpack it, and follow the installation instructions. For upgrades, select the same installation location.

💾 Download Installer

💻 Developer Changes

v4.0.0 is versioned as a major release due to breaking changes:

  • The internal nodes API has been refactored to provide a stable public API. Node authors should review the migration guide.
  • The internal graph execution engine is drastically simplified, resulting in more efficient and performant processing. This carries on from the changes in v3.6.0 in which graphs are no longer stored in the database.

Contributing

As a community-supported project, we rely on volunteers and enthusiasts for continued innovation and polish. Everything from minor documentation fixes to major feature additions are welcome. To get started as a contributor, please refer to How to Contribute or reach out in #dev-chat on Discord!

What's Changed

Full Changelog: v4.0.0rc1...v4.0.0rc2

Contributors

lstein, dunkeroni, and 7 other contributors
Assets 3
Loading